Edge but not Least: Cross-View Graph Pooling

Graph neural networks have emerged as a powerful representation learning model for undertaking various graph prediction tasks. Various pooling methods been developed to coarsen an input into succinct graph-level through aggregating node embeddings obtained via convolution. However, because most are heavily node-centric, they fail fully leverage the crucial information contained in structure. This paper presents cross-view method (Co-Pooling) that explicitly exploits substructures representations. Co-Pooling is designed fuse pooled representations from both view and edge view. Through interaction, edge-view node-view mutually reinforce each other learn informative Extensive experiments on one synthetic 15 real-world datasets validate effectiveness of our method. Our results analysis show (1) able yield promising over graphs with types attributes, (2) can achieve superior performance state-of-the-art classification regression